Decompression techniques for processing compressed data suitable for artificial neural networks

A non-transitory computer readable medium storing code executable by a processor to perform a method for decompressing compressed data comprising:

• receiving as input at least a portion of an N element data structure, the N element data structure being a data structure having N number of elements,
• where the N element data structure was generated by compressing an original M element data structure, the original M element data structure being a data structure having M number of elements,
• where each element in the N element data structure holds a value, and
• where N<M;
• receiving as input at least a portion of metadata associated with the N element data structure, the metadata specifying for each value in the N element data structure one of the M elements of the original M element data structure in which the value originated;
• identifying at least a portion of a destination M element data structure, the destination M element data structure being a destination data structure having M number of elements;
• and
• executing at least one instruction to decompress at least a portion of the N element data structure into the destination M element data structure by:
• identifying one or more values of the N element data structure to be written to the destination M element data structure,
• determining, from the input metadata for each of the one or more values of the N element data structure, in which one of the M elements of the original M element data structure the value of the N element data structure originated, and
• writing each of the one or more values of the N element data structure to an element of the destination M element data structure that corresponds to the one of the M elements of the original M element data structure in which the value of the N element data structure originated.

A system for decompressing compressed data comprising:

• a memory;
• and
• at least one processor for:
• receiving as input at least a portion of an N element data structure, the N element data structure being a data structure having N number of elements,
• where the N element data structure was generated by compressing an original M element data structure, the original M element data structure being a data structure having M number of elements,
• where each element in the N element data structure holds a value, and where N<M;
• receiving as input at least a portion of metadata associated with the N element data structure, the metadata specifying for each value in the N element data structure one of the M elements of the original M element data structure in which the value originated;
• identifying at least a portion of a destination M element data structure, the destination M element data structure being a destination data structure having M number of elements;
• and
• executing at least one instruction to decompress at least a portion of the N element data structure into the destination M element data structure by:
• identifying one or more values of the N element data structure to be written to the destination M element data structure,
• determining, from the input metadata for each of the one or more values of the N element data structure, in which one of the M elements of the original M element data structure the value of the N element data structure originated, and
• writing each of the one or more values of the N element data structure to an element of the destination M element data structure that corresponds to the one of the M elements of the original M element data structure in which the value of the N element data structure originated.

A method for partially decompressing compressed data, comprising:

• receiving as input at least a portion of an N element data structure, the N element data structure being a data structure having N number of elements,
• where the N element data structure was generated by compressing an original M element data structure, the original M element data structure being a data structure having M number of elements,
• where each element in the N element data structure holds a value, and
• where N<M;
• receiving as input at least a portion of metadata associated with the N element data structure, the metadata specifying for each value in the N element data structure one of the M elements of the original M element data structure in which the value originated;
• identifying at least two destination R element data structures, the at least two destination R element data structures each being a destination data structure having R number of elements,
• where N<=R<M, and
• where each of the at least two destination R element data structures represents a different portion of a destination M element data structure;
• and
• executing at least one instruction to decompress the N element data structure to the at least two destination R element data structures by:
• identifying at least two portions of the input N element data structure to be stored in the at least two destination R element data structures,
• determining, from the input metadata for each of the at least two portions of the input N element data structure, in which one of the M elements of the original M element data structure the portion of the input N element data structure originated, and
• storing each of the at least two portions of the input N element data structure in one of the at least two destination R element data structures that represents the portion of the destination M element data structure which corresponds with the one of the elements of the original M element data structure in which the portion of the input N element data structure originated.